221 | 🔙 🤔 Back to Basics: How to get better responses (shots + priming)

In today’s Brainyacts:

1. How safe is GPT-4o (how OpenAI tests it)

2. Back to Basics Prompting: Shots & Priming video

3. Google launches Android voice assistant and other AI model news

4. “Compressed copies” copyright complaint survives Motion to Dismiss and more news you can use

   👋 to all subscribers!

To read previous editions, click here.

Lead Memo

🦺⛑️ How OpenAI’s tries to ensure that GPT-4o is safe

The GPT-4o System Card is a detailed report that outlines the safety measures and risk assessments undertaken before releasing the GPT-4o model, particularly its new audio capabilities. Here’s a plain-language summary and key points for general users:

Summary:

1. Safety and Risk Mitigations: Before releasing GPT-4o, OpenAI conducted extensive safety evaluations, including external testing (red teaming) and risk assessments based on their Preparedness Framework. The model was assessed for potential risks, especially concerning its new audio features, and safeguards were put in place to minimize these risks.

2. Key Risks Evaluated:

   • Unauthorized Voice Generation: Ensuring the model doesn't generate voices that could be used maliciously.

   • Speaker Identification: Limiting the model's ability to identify people based on voice inputs to protect privacy.

   • Ungrounded Inference & Sensitive Trait Attribution: Preventing the model from making assumptions about people (e.g., intelligence or traits) based on voice alone.

   • Disallowed Audio Content: Blocking the generation of harmful, violent, or explicit audio content.

   • Erotic & Violent Speech: Ensuring the model does not generate inappropriate content.

3. Performance and Preparedness:

   • Low to Medium Risk: GPT-4o’s overall risk is categorized as medium, with specific risks like persuasion (ability to influence users) assessed as borderline medium. Other risks like cybersecurity and model autonomy were rated low.

   • Improvements Over Previous Models: GPT-4o offers better performance in non-English languages, faster responses, and enhanced vision and audio understanding compared to earlier models.

4. Societal Impacts:

   • GPT-4o's human-like voice capabilities could lead to users forming emotional attachments or anthropomorphizing the AI, which might affect social norms and human interactions.

   • Potential positive impacts include accelerating scientific research and improving access to health information.

   • Efforts are being made to ensure the model is fair and performs well across various languages, especially those that are historically underrepresented.

Spotlight

🔙 🤔 Back to Basics: How to get better responses (shots + priming)

So, I just made this quick video about two super useful tricks for getting AI to actually understand what you're asking: "shots" and "priming". Trust me, it's not as complicated as it sounds!

Shots are basically examples you give the AI before asking it to do something. No examples? That's "zero-shot". One example? You guessed it, "one-shot". A few examples? You got it, "few-shot" or "multi-shot".

Priming is just setting the scene for the AI. It's like giving it some context before you ask your main question. It helps the AI get what you're talking about and give you better answers.

Here's What You'll Pick Up from My Video:

• Why just throwing a question at AI usually gets you a meh response

• How AI is totally different from Google when it comes to understanding you

• A super simple way to use those "shot" techniques I mentioned

• My step-by-step method for priming AI (it's easier than you think!)

• A real-life example using a sci-fi story (yeah, I'm a bit of a nerd)

• Cool tips for getting AI to really think about your questions

Whether you're new to GenAI or you've been at it for a while, these tricks will seriously up your game. You'll get way better responses, and it'll feel like the AI actually gets what you're after.

So, take a few minutes to watch the video. I promise it'll make your chats with AI way more useful (and probably a lot more fun too)!

AI Model Notables

► Just announced: Google launches its new voice assistant on Andriod phones

• Gemini Live is now available for interactive, mobile conversations with Gemini.

• Hands-free use allows continued interaction even when the app runs in the background or the phone is locked.

• Rolling out in English to Gemini Advanced subscribers on Android; coming soon to iOS and more languages.

• 10 new voices have been added for personalized tone and style choices.

► This AI startup focuses on Personal Injury plaintiffs and their lawyers: AI-generated demand letters can save up to 15 hours per case and result in settlements as much as 30% higher than plaintiffs might otherwise get.

► Free ChatGPT users can now use image generator DALL-E for free; but you are limited to two images per day

► Melior AI, a legal document AI platform has just launched a crowdfunding campaign on republic.com/melior, offering a rare chance for investors to get in on the ground floor of this legal tech revolution. THIS IS NOT FINANCIAL ADVICE OR A RECOMMENDATION.

► Amazon's Alexa division lost $10 billion in 2022 alone, prompting layoffs and a pivot to generative AI to revitalize the smart assistant.

► Universal Music and Meta announce ‘Expanded Global Agreement’ for AI, monetization and more

► OpenAI is worried that ChatGPT-4o users are developing feelings for the chatbot

News You Can Use:

➭ Judge allows “compressed copies” theory to advance in text-to-image generation class action suit. Read opinion HERE:

• Judge Orrick has ruled that “the allegations of induced copyright infringement are sufficient” enough for the case to move forward into a discovery phase which will determine if these companies violated copyright laws.

• During the discovery phase, the artist's lawyers will be given access to internal documents from the indicted AI companies and will reveal details about their training datasets, mechanisms, and inner workings.

➭ Replika CEO Eugenia Kuyda says it’s okay if we end up marrying AI chatbots

➭ New York State Attorney General releases report on AI outlining areas of interest for policy and rulemaking.

Who is the author, Josh Kubicki?

Some of you know me. Others do not. Here is a short intro. I am a lawyer, entrepreneur, and teacher. I have transformed legal practices and built multi-million dollar businesses. Not a theorist, I am an applied researcher and former Chief Strategy Officer, recognized by Fast Company and Bloomberg Law for my unique work. Through this newsletter, I offer you pragmatic insights into leveraging AI to inform and improve your daily life in legal services.